The thermal compensation technology for 3-axis machine tools is quite mature, while 5-axis machine tools are more complicated due to the addition of two rotary axes to the overall thermal deformation. In this study, we measured the thermal displacement of 5-axis machine tools by Eddy-Current Displacement Measurement Equipment, and simulated the actual cutting conditions by using the analog cutting method to get closer to the actual displacement situation. In contrast to the previous literature, where most measurements were made only in axial static conditions or only in specific axes without real 5-axis simultaneous motion, the causes and effects of dislocations in linear and rotary axes are investigated by means of the measured data. The thermal compensation model utilizes the multiple linear regression method and establishes three kinds of learning models, which are single day learning file (5-axis stationary), single day learning file (5-axis simultaneous motion) and learning different days file, and the results of the validation show that the learning different days file modeling is the optimal way and only requires three temperature points. This model can be used to predict the general cutting conditions, which can effectively improve the Z-axis, which has the most severe deformation, from an average maximum displacement of 28.7 μm to 9.2 μm, and it can be used to predict the spindle stoppage and spindle speed change in low temperature environments, which can also effectively reduce the average error of the X-axis, Y-axis, and Z-axis to less than 3.1 μm.